Mainak Mallick

Machine Learning Engineer and Full Stack Developer with a unique blend of research and industry experience, currently pursuing a Master's in Computational Science and Engineering at Georgia Tech. Specializing in developing scalable AI solutions, with a focus on few-shot learning, digital twins, and real-time systems.

At Georgia Tech, pioneering research in predictive maintenance using meta-learning (MAML) and digital twins, achieving 86% accuracy in fault detection. Engineering high-performance full-stack applications with Vue.js, Flask, and Redis, while implementing efficient CI/CD pipelines using Docker and AWS.

Previously at HSBC, led critical software initiatives developing microservices and full-stack applications impacting £3M+ in portfolio assets. Recent projects include benchmarking LLaMA models with ReFT and IA3 techniques, and implementing real-time 3D prediction systems using YOLO and CLIP, demonstrating expertise across the full technical stack.

Mainak Mallick profile photo

Professional Experience

Predictive maintenance of Robotic Hand with Digital Twin and MAML algorithm

Graduate Research Assistant
Dec 2024 - Present
Georgia Tech (Supervisor – Dr. Seung-Kyum Choi)
  • Generated synthetic sensor data using Isaac Sim to simulate real-world movement of a KUKA LBR robot arm.
  • Developed a few shot learning pipeline using the MAML algorithm for fault classification in difference joints.
  • Achieved over 86.4% testing accuracy by optimizing the model with a minimal dataset.
Python Icon Python PyTorch Icon PyTorch Isaac Sim Icon Isaac Sim

Full Stack Development

Graduate Research Assistant
Aug 2024 - Dec 2024
Georgia Tech (Supervisor – Dr. Seung-Kyum Choi)
  • Built responsive UIs using Vue.js and E-Charts to visualize real-time sensor data, reducing access time by 20%.
  • Engineered REST APIs with Flask, integrated WebSocket and Redis, cutting data latency by 28% and supporting 30% more concurrent users.
  • Streamlined CI/CD pipelines and deployments using Docker, GitHub Actions, and AWS S3, improving deployment speed by 40% and reducing manual errors by 50%.
Vue.js Icon Vue.js Flask Icon Flask Redis Icon Redis Docker Icon Docker AWS Icon AWS

Managed Collections and Recoveries Decision systems

Software Engineer
July 2022 - Aug 2024
HSBC Bank
  • Built and maintained full-stack web apps for risk analysis tools using React, Node.js, and MongoDB.
  • Developed REST APIs and microservices with Spring Boot, Java, and PostgreSQL for data integration.
  • Deployed 24+ business critical code changes in SAS EG and SAS ID env., impacting over £3 million in assets.
React Icon React Node.js Icon Node.js MongoDB Icon MongoDB Spring Boot Icon Spring PostgreSQL Icon PostgreSQL

Research

Ensemble Meta-Learning Research Visualization

Ensemble-Based Model-Agnostic Meta-Learning...

Mainak Mallick, Young-Dae Shim, Seung-Kyum Choi
MDPI, Sensors, 2024

  • Developed an ensemble meta-learning framework with digital twins for real-time robotic arm fault classification.
  • Utilized Latin Hypercube Sampling (LHS) for enhanced generalization and stability in cross-domain fault detection.
  • Demonstrated superior performance over ANIL, ProtoNet, and Reptile in classification accuracy on synthetic vibration data.
Honeycomb Panel Optimization Visualization

Genetic Algorithm-Based Design Optimization...

Mainak Mallick, A. Chakrabarty, N. Khutia
Materials Today, 2022

  • Optimized honeycomb sandwich panels for improved impact energy absorption via NSGA-II.
  • Integrated MATLAB-ABAQUS with Python scripting to optimize core cell count and wall thickness.
  • Achieved significant improvements in panel performance metrics through multi-objective optimization.

Projects

LLaMA Fine-Tuning Project Snippet

Fine-Tuning LLaMA 3B for Code Completion Jan 2025 - Present

  • Comparing ReFT vs IA3 for parameter-efficient fine-tuning of LLAMA 3B on HumanEval.
  • Analyzing ReFT's representation edits vs IA3's activation scaling for program synthesis.
  • Benchmarking ReFT's efficiency gains against IA3's lightweight adaptation on coding tasks.
Volumetric Occupancy Prediction Project Visualization

Volumetric Occupancy Prediction & Labeling Aug 2024 - Dec 2024

  • Developed ISO model-based framework for volumetric occupancy prediction indoors.
  • Integrated YOLO for real-time object detection and segmentation.
  • Employed CLIP for semantic labeling, enabling contextual understanding of scenes.
AI Research Recommender Project Interface

AI-Powered Research Paper Recommender Jun 2022 - Aug 2022

  • Built a recommender using React, Node.js, MongoDB, OpenAI API.
  • Integrated arXiv/Semantic Scholar APIs for paper fetching and summary generation.
  • Utilized OpenAI for AI-driven, personalized paper suggestions based on user input.
Automated Defect Detection Project Visualization

Automated Defect Detection in Manufacturing using Self-Supervised Anomaly Localization May 2025 - Aug 2025

  • Developed a self-supervised learning pipeline using contrastive methods (e.g., SimCLR) on PyTorch to learn robust visual features from unlabeled industrial image data, significantly reducing manual annotation efforts.
  • Engineered and evaluated an anomaly localization module (e.g., PatchCore) achieving an F1-score of 92.5% in identifying and segmenting diverse defects on the MVTec AD benchmark.
Controllable Artistic Style Transfer Project Visualization

Controllable Artistic Style Transfer with Stable Diffusion via Cross-Attention Modulation Sep 2025 - Dec 2025

  • Investigated cross-attention map manipulation within Stable Diffusion (SDXL) and fine-tuned using Low-Rank Adaptation (LoRA) for precise artistic style transfer from complex textual prompts.
  • Demonstrated 15% improved style adherence (CLIP/FID metrics) over baseline prompting techniques and deployed an interactive Gradio demo for personalized, controllable image generation.
AI Agent Latex Editor Project Visualization

AI Agent based Latex Editor Web App Aug 2024 - Dec 2024

  • Developed an AI-Agent integrated Latex Editor using Vue, Node.js, MongoDB, and OpenAI API to automatically edit the Latex Script based on user input and search queries.
  • Integrated external APIs (arXiv and Semantic Scholar) to fetch research papers and used OpenAI API for generating personalized summaries and inputs.